pyannote.audio

pyannote.audio is the open-source standard for speaker diarization — the “who spoke when” problem — a PyTorch toolkit with more than 10,600 stars on GitHub and over a billion downloads of its models.

Overview

Transcripts tell you what was said; diarization tells you who said it, and when. pyannote.audio provides the neural building blocks for that job: speech activity detection to find where speech happens, speaker change detection to find where turns switch, overlapped-speech detection for the messy moments when people talk over each other, and speaker embeddings that identify and cluster voices. These components assemble into a complete diarization pipeline whose pretrained models have effectively become the market standard, used by transcription platforms, meeting tools, dubbing studios, and call-analytics products worldwide. Everything is built on PyTorch, models are shared through Hugging Face, and the toolkit doubles as a training framework if you want to adapt the models to your own audio. The same team now offers the Precision API through pyannote.ai — the same interface with higher-accuracy production models — for teams that outgrow the open-source deployment.

Key Features

• Speaker diarization pipeline – production-ready “who spoke when” out of the box.
• Neural building blocks – voice activity, speaker change, and overlap detection as separate components.
• Speaker embeddings – identify, verify, and cluster speakers across recordings.
• Pretrained models – state-of-the-art checkpoints distributed via Hugging Face.
• Trainable framework – fine-tune and train components on your own datasets with PyTorch.
• Pairs with any STT – combine with Whisper or any transcription engine for attributed transcripts.
• Industry standard – 1B+ downloads and adoption across transcription, meetings, and media.

💳 Pricing

pyannote.audio is free and open source under the MIT license — start there at no cost. For production scale, pyannoteAI offers the hosted Precision API with higher-accuracy models on paid plans, or deployment on your own infrastructure.

Who Is It For

Speech and voice-AI developers, transcription and meeting platforms, media and dubbing teams, and researchers who need reliable speaker attribution in real-world, multi-speaker audio.